PRATIK/SYS

RESEARCH ENGINEER · TOKYO

I build software systems that have to reason, execute, and survive reality.

My work sits at the boundary of AI systems, model-based engineering, software automation, and agent reliability — from Linux networking and large-scale repository pipelines to RAG, LLM agents, and formal verification.

Inspect selected systems ↓
31%development lead-time reduction
10K+heterogeneous documents benchmarked
93%Recall@10 in retrieval study
86%network adapter reboot-time reduction

01 / SELECTED SYSTEMS

Projects shown as engineering decisions, not feature lists.

A1
AI SYSTEMS · RETRIEVAL · EVALUATION

Agentic RAG benchmark for heterogeneous technical corpora

Built an end-to-end evaluation system over PDFs, HTML, DOCX, CSV and tabular data to test where retrieval architecture and agentic routing actually improve answer quality — and where they only add latency.

query→router→BM25 + dense→RRF→cross-encoder→tools / generator
Benchmark500 queries

Factual, multi-hop, cross-document, tabular, and unanswerable cases.

Retrieval76 → 93%

Recall@10 improvement across controlled retrieval experiments.

Ranking0.68 → 0.86

NDCG@10 improvement using hybrid retrieval and reranking.

System trade-off2.3×

Lower latency for simpler RAG on workloads where agent routing added no quality.

What mattered

Separating retrieval quality from generation quality. Treating latency and token cost as first-class evaluation dimensions. Designing tests that can expose when “more agentic” is not “better.”

Python · BM25 · dense embeddings · RRF · cross-encoders · vLLM · llama.cpp · RAG evals

A2
MODEL-BASED ENGINEERING · AUTOMATION

Automation platform for safety-critical Simulink development

Engineered tooling that generates, configures, tests, documents, and integrates Simulink models with legacy C++ — turning repetitive model engineering into a reproducible software pipeline.

spec / inputs→model generation→C++ integration→test harness→report
Impact31%

Lead-time reduction for model development workflows in power protection and elevator applications.

Engineering focusLegacy → model

Integrated object-oriented C++ into Simulink workflows without forcing source-level redesign.

System qualities

Deterministic generation, reusable component libraries, batch test execution, automatic reporting, and practical adoption inside existing engineering processes.

MATLAB · Simulink · Embedded Coder · C++ · Python · PySide6

A3
DATA PIPELINES · PARALLEL PROCESSING · ML

Repository intelligence pipeline across 1,000+ codebases

Built a per-user analysis application for repository-scale engineering data. The pipeline ingests repositories, parses them in parallel, persists intermediate artifacts, computes features incrementally, aggregates results, executes ML stages, and produces reports.

ingestion→parallel parse→features→persistent artifacts→aggregation→ML→reporting

Why it is a systems project

Asynchronous jobs, caching and incremental computation, persistent storage, failure recovery, observability, and deployment mattered as much as the ML itself.

Python · Panel · multiprocessing · persistent cache · ML pipelines

A4
SYSTEMS SOFTWARE · LINUX NETWORKING

Storage-controller networking and live-maintenance tooling

Worked below the application layer on storage platform software: Linux networking, controller debug tooling, live patching, and restart behavior for network adapters.

Recovery−86%

Reduced network-adapter OS reboot time through systems-level optimization.

I/O+100%

Improved I/O performance in storage networking work.

Foundation

This is the layer that shaped how I approach AI systems today: resource limits, failure modes, latency, observability, and interfaces are not implementation details.

C/C++ · Linux · kernel / networking · NVMe-TCP · iSCSI

02 / CURRENT RESEARCH

How do you verify an agent whose plan can change every time?

“Verification should follow the state and dependencies an agent actually touches — not re-check the entire world after every change.”

I am exploring selective verification for non-deterministic agents: planner-declared dependencies, runtime provenance, state drift, incremental re-verification, and formal constraints.

PLAN
DEPENDENCIES
RUNTIME
PROVENANCE
STATE
DRIFT
SELECTIVE
VERIFY
Z3 / SMTTLA+runtime verificationprogram slicingprovenanceagent orchestrationstateful workflows

03 / TRAJECTORY

One through-line: making complex systems more inspectable and automatable.

Systems software

Storage controllers, Linux networking, performance, debugging and live-maintenance tooling.

C/C++ · Linux
Repository-scale analysis

Parallel parsing, persistent intermediate artifacts, feature extraction, ML pipelines and reporting.

Python · data systems
Model-based development automation

Simulink automation, C++ integration, test generation, engineering GUIs, documentation workflows.

MATLAB · Simulink
AI systems & agent reliability

RAG, tool-using agents, retrieval evaluation, inference benchmarking, verification and state-aware re-checking.

LLMs · formal methods

04 / ABOUT

Research depth with an implementation bias.

I am a researcher and software engineer at Hitachi in Japan. My background crosses control, embedded and systems software, model-based development, and modern AI engineering. I prefer problems where the hard part is not a single model or algorithm, but the behavior of the whole system.

That means I care about retrieval quality and latency; agent reasoning and runtime state; model generation and testability; correctness and deployment. I build prototypes to answer technical questions, then turn the useful ones into repeatable engineering systems.

M.Tech · Systems & Control EngineeringIIT Bombay

Autonomous vehicle decision-making under uncertainty; MDP/POMDP, MCTS, reinforcement learning.

B.Tech · Mechanical EngineeringIIT Indore

Engineering foundation with early robotics and technical leadership experience.

AI / LLM

RAG, BM25, dense retrieval, RRF, reranking, agents, MCP, evals, vLLM, llama.cpp

Software

Python, C++, MATLAB, JavaScript, FastAPI, async systems, caching, queues

Systems

Linux, networking, Docker, CI, parallel processing, persistent pipelines

Robotics / control

MPC, MCTS, POMDPs, EKF, ROS 2, path tracking, simulation

WHAT I AM OPTIMIZING FOR

AI engineering roles where reasoning systems meet real software constraints.

Especially work involving agent infrastructure, retrieval systems, verification, developer tooling, robotics software, or research engineering.

Copied.